feat(backend): write changed-path Bloom filters to commit-graph

Writes --changed-paths when building the commit-graph, which adds Bloom
filters that let git cheaply skip commits that did not touch a given
path. This dramatically accelerates `git log -- <path>` and modestly
helps `git blame` on large repos.

Existing repos that already have a Bloom-less commit-graph get a
one-time `--split=replace` rewrite on their next fetch, gated on a new
`commitGraphChangedPathsBackfilledAt` timestamp stored in repo metadata.
Subsequent fetches do a cheap incremental write.

Also moves the `writeCommitGraph` call out of `cloneRepository` and into
`RepoIndexManager.indexRepository` so clone and fetch paths handle the
commit-graph symmetrically. Drops `--write-commit-graph` from the fetch
invocation since that flag does not honor `--changed-paths`.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brendan-kellam

packages/backend/src/git.ts

@@ -105,12 +105,6 @@ export const cloneRepository = async (
             keys: ["remote.origin.url"],
             signal,
         });
-
-        // @note: operations that need to iterate over a lot of commits (e.g., rev-list --count)
-        // can be slow on larger repositories. Commit graphs are a acceleration structure that
-        // speed up these operations.
-        // @see: https://git-scm.com/docs/commit-graph
-        await writeCommitGraph({ path, signal });
     } catch (error: unknown) {
         const baseLog = `Failed to clone repository: ${path}`;
 
@@ -151,9 +145,6 @@ export const fetchRepository = async (
             "+refs/heads/*:refs/heads/*",
             "--prune",
             "--progress",
-            // On fetch, ensure the commit graph is up to date.
-            // @see: https://git-scm.com/docs/commit-graph
-            "--write-commit-graph"
         ]);
 
         // Update HEAD to match the remote's default branch. This handles the case where the remote's
@@ -490,20 +481,42 @@ export const isRepoEmpty = async ({
  * Writes or updates the commit-graph file for the repository.
  * This pre-computes commit metadata to speed up operations like
  * rev-list --count, log, and merge-base.
+ *
+ * Also writes changed-path Bloom filters (--changed-paths), which let git
+ * quickly skip commits that didn't touch a given path. This accelerates
+ * `git log -- <path>` dramatically and `git blame` modestly on large repos.
+ *
+ * For incremental writes (the default), Bloom filters are only computed for
+ * commits being added in this write. Repos that already have a Bloom-less
+ * commit-graph from a prior version need a one-time `forceBackfill: true`
+ * write to backfill filters for their historical commits.
+ *
+ * @see: https://git-scm.com/docs/commit-graph
  */
 export const writeCommitGraph = async ({
     path,
+    forceBackfill,
     onProgress,
     signal,
 }: {
     path: string,
+    forceBackfill?: boolean,
     onProgress?: onProgressFn,
     signal?: AbortSignal,
 }): Promise<void> => {
     const git = createGitClientForPath(path, onProgress, signal);
 
     try {
-        await git.raw(['commit-graph', 'write', '--reachable']);
+        const args = ['commit-graph', 'write', '--reachable', '--changed-paths'];
+        if (forceBackfill) {
+            // --split=replace consolidates any existing layers into a single new layer,
+            // which forces git to recompute Bloom filters for every commit (not just commits
+            // added since the last write). Used for the one-time migration of repos that have
+            // a Bloom-less commit-graph from before --changed-paths was enabled.
+            // @see: https://git-scm.com/docs/git-commit-graph#Documentation/git-commit-graph.txt-write
+            args.push('--split=replace');
+        }
+        await git.raw(args);
     } catch (error) {
         // Don't throw an exception here since this is just a performance optimization.
         logger.debug(`Failed to write commit-graph for ${path}:`, error);

packages/backend/src/repoIndexManager.test.ts

@@ -51,6 +51,7 @@ vi.mock('./git.js', () => ({
     isRepoEmpty: vi.fn().mockResolvedValue(false),
     unsetGitConfig: vi.fn(),
     upsertGitConfig: vi.fn(),
+    writeCommitGraph: vi.fn(),
 }));
 
 vi.mock('./zoekt.js', () => ({
@@ -178,6 +179,8 @@ const createMockSettings = (): Settings => ({
     enablePublicAccess: false,
     experiment_repoDrivenPermissionSyncIntervalMs: 1000 * 60 * 60 * 24,
     experiment_userDrivenPermissionSyncIntervalMs: 1000 * 60 * 60 * 24,
+    repoDrivenPermissionSyncIntervalMs: 1000 * 60 * 60 * 24,
+    userDrivenPermissionSyncIntervalMs: 1000 * 60 * 60 * 24,
     maxAccountPermissionSyncJobConcurrency: 8,
     maxRepoPermissionSyncJobConcurrency: 8,
 });

packages/backend/src/repoIndexManager.ts

@@ -8,7 +8,7 @@ import { Redis } from 'ioredis';
 import micromatch from 'micromatch';
 import Redlock, { ExecutionError } from 'redlock';
 import { INDEX_CACHE_DIR, REPOS_CACHE_DIR, WORKER_STOP_GRACEFUL_TIMEOUT_MS } from './constants.js';
-import { cloneRepository, fetchRepository, getBranches, getCommitHashForRefName, getLatestCommitTimestamp, getLocalDefaultBranch, getTags, isPathAValidGitRepoRoot, isRepoEmpty, unsetGitConfig, upsertGitConfig } from './git.js';
+import { cloneRepository, fetchRepository, getBranches, getCommitHashForRefName, getLatestCommitTimestamp, getLocalDefaultBranch, getTags, isPathAValidGitRepoRoot, isRepoEmpty, unsetGitConfig, upsertGitConfig, writeCommitGraph } from './git.js';
 import { captureEvent } from './posthog.js';
 import { PromClient } from './promClient.js';
 import { RepoWithConnections, Settings } from "./types.js";
@@ -396,6 +396,19 @@ export class RepoIndexManager {
             const fetchDuration_s = durationMs / 1000;
 
             logger.debug(`Fetched ${repo.name} (id: ${repo.id}) in ${fetchDuration_s}s`);
+
+            // Update the commit-graph after fetch. Force a full backfill the first time we
+            // see this repo after the --changed-paths rollout, so historical commits get
+            // Bloom filters. Subsequent fetches do a cheap incremental write.
+            const needsBackfill = !metadata.commitGraphChangedPathsBackfilledAt;
+            if (needsBackfill) {
+                logger.debug(`Backfilling changed-path Bloom filters for ${repo.name} (id: ${repo.id})...`);
+            }
+            await writeCommitGraph({
+                path: repoPath,
+                forceBackfill: needsBackfill,
+                signal,
+            });
         } else if (!isReadOnly) {
             logger.debug(`Cloning ${repo.name} (id: ${repo.id})...`);
 
@@ -411,6 +424,27 @@ export class RepoIndexManager {
             const cloneDuration_s = durationMs / 1000;
 
             logger.debug(`Cloned ${repo.name} (id: ${repo.id}) in ${cloneDuration_s}s`);
+
+            // Write the commit-graph for the freshly cloned repo.
+            await writeCommitGraph({
+                path: repoPath,
+                signal,
+            });
+        }
+
+        // Record that this repo's commit-graph now includes changed-path Bloom filters
+        // for its full history (either freshly written during clone, or backfilled above
+        // during fetch).
+        if (!isReadOnly && !metadata.commitGraphChangedPathsBackfilledAt) {
+            await this.db.repo.update({
+                where: { id: repo.id },
+                data: {
+                    metadata: {
+                        ...metadata,
+                        commitGraphChangedPathsBackfilledAt: new Date().toISOString(),
+                    } satisfies RepoMetadata,
+                },
+            });
         }
 
         // Regardless of clone or fetch, always upsert the git config for the repo.

packages/shared/src/types.ts

@@ -32,6 +32,14 @@ export const repoMetadataSchema = z.object({
      */
     indexedRevisions: z.array(z.string()).optional(),
 
+    /**
+     * Timestamp of when changed-path Bloom filters were written into the
+     * commit-graph for this repo's full history. Undefined means the one-time
+     * backfill has not yet run, so historical commits still lack Bloom filters.
+     * @see writeCommitGraph in packages/backend/src/git.ts
+     */
+    commitGraphChangedPathsBackfilledAt: z.string().datetime().optional(),
+
     /**
      * Code host specific metadata, keyed by code host type.
      */

packages/web/src/app/(app)/search/components/codePreviewPanel/codePreview.tsx

@@ -112,7 +112,7 @@ export const CodePreview = ({
 
     return (
         <div className="flex flex-col h-full">
-            <div className="flex flex-row bg-accent items-center justify-between pr-3 py-0.5 mt-7">
+            <div className="flex flex-row bg-accent items-center justify-between pr-3 py-[3px] mt-7">
 
                 {/* Gutter icon */}
                 <div className="flex flex-row">